The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill manages its dependencies via npm, as defined in the assets/package.json file. It relies on standard libraries for browser automation and image processing, such as playwright-core, pixelmatch, and pngjs.
[COMMAND_EXECUTION]: The skill documentation instructs the user to install dependencies and run the route-shots.mjs driver script. This script performs necessary file system operations, including directory creation and reading/writing image files.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface where the agent processes visual data (screenshots) from a web application.
Ingestion points: Visual content from screenshots and metadata in MANIFEST.json processed in SKILL.md and route-shots.mjs.
Boundary markers: Structured rubrics in SKILL.md and automated pixel-diffing are used to guide the agent and ensure objective review.
Capability inventory: The script has permissions to interact with the local file system and automate browser tasks on localhost.
Sanitization: Deterministic checks for color contrast and layout consistency provide objective metrics that supplement and validate the agent's analysis.
[DYNAMIC_EXECUTION]: The route-shots.mjs script uses page.evaluate() to execute the DESIGN_AUDIT function within the web page context. This is a standard method for performing in-page analysis during browser automation and involves executing locally-defined logic.

visual-qa